CN104462762A - Fuzzy fault classification method of electric transmission line - Google Patents

Abstract

A fuzzy fault classification method of an electric transmission line includes the first step of determining the time of occurrence of a fault, the second step of computing fault input vectors, the third step of constructing fuzzy support vector machine FSVM dichotomy devices, the fourth step of training and optimizing the FSVM dichotomy devices, the fifth step of constructing a banding subsection subordinating degree function of a FSVM higher space, the sixth step of enabling the fault input vectors to be input into each FSVM dichotomy device to obtain a preliminary classification label, a decision function value and an initial subordinating degree of each FSVM dichotomy device, the seventh step of constructing and training a support vector regression (SVR), the eighth step of sending the decision function values and initial subordinating degrees into the SVR to obtain a final fault subordinating degree of a fault sample, and the ninth step of judging the final fault type according to the final subordinating degree. According to the fuzzy fault classification method of the electric transmission line, the fuzzy subordinating degree function is introduced, and therefore influences of noise points and isolated points on a SVM hyperplane structure are reduced; the SVR is adopted to perform correction on the preliminary classification labels obtained by the FSVM, the fault classification label is obtained accurately through fuzzification processing, regressive optimization processing and the like, and therefore the accuracy and fault tolerance for fault classification of the electric transmission line are greatly improved.

Description

A kind of fuzzy fault sorting technique of transmission line of electricity

Technical field

The present invention relates to power system transmission line trouble hunting field, particularly a kind of fuzzy fault sorting technique of transmission line of electricity.

Background technology

Transmission line fault classification is very important for the normal operation of electric system.Feature at present for transmission line malfunction classification is more, if mainly take Fourier transform to process original electric current, voltage signal, frequency domain character [1], temporal signatures [2], time-frequency characteristics [3] three major types characteristic quantity can be summarized as according to the character of characteristic quantity.

Document [4] utilizes (the Empirical Mode Decomposition of empirical mode decomposition in Hilbert-Huang transform, EMD) decomposed signal localization characteristic, obtain the singular value entropy of sample signal, carry out the failure modes of transmission line of electricity in this, as characteristic quantity.

Document [5] utilizes EMD to construct the Sample Entropy of sample signal, as characteristic quantity, then adopts extreme learning machine to carry out transmission line malfunction type identification.

Document [6] is a patent, it relates to a kind of resonant earthed system classification line selection method for single-phase earth fault, its way is all waveforms of a power frequency of circuit after read failure, EMD decomposition is carried out to the fault zero-sequence current of half power frequency cycle, obtain Hilbert time-frequency spectrum and the Hilbert marginal spectrum of each bar circuit transient zero-sequence current, calculate Hilbert time-frequency entropy, and adopt support vector machine to classify to different faults type.But this patent, only for resonant earthed system singlephase earth fault, can not distinguish other type fault, only make use of transient zero-sequence current signal, and adopts standard supporting vector machine model, does not have Fuzzy Processing ability.

Document [7] adopts Fourier transform, and during by extracting fault, the signal characteristic quantity of each sequence fundamental current, voltage, in conjunction with least square method supporting vector machine for the training of transient signal in phase selection and classification, exports the classification results of+1 or-1.But it adopts traditional Fourier transform to obtain each sequence fundamental current and the voltage signal input quantity as support vector machine.

Document [8] application wavelet transformation technique extracts the residual voltage low frequency signal energy of reflection earth fault feature, and application message entropy extracts the wavelet singular entropy of three-phase voltage.With the wavelet singular entropy of residual voltage low frequency energy and three-phase voltage for input feature vector amount, with separate A, B, C and ground G for output quantity, establish SVM fault type recognition network that four inputs four export to identify fault type.

Document [9] proposes the high voltage transmission line fault type recognition new method based on fractal theory.According to after fault to the fractal dimension computation and analysis of each phase transient state component electric current and zero sequence transient current, extract and define fault type recognition criterion.The method has good validity to transmission line of electricity.

Document [10] uses the three-phase current of faulty line and zero-sequence current as the input of RBF nerve network, utilize the feature of every phase current to determine whether this phase has fault, utilize zero-sequence current to judge whether and earth fault occurs, but when this neural network is for identifying the failure condition of the adjacent lines of faulty line or parallel circuit, be as easy as rolling off a logly identified as earth fault.In the display of two classification result visualization, document [11] in some application to two-category data classification results visualization requirement, propose two-category data classification results visualized algorithm.This algorithm is at unsupervised self organizing neural network (Self-Organizing Map, on the basis of SOM) visualization function, in conjunction with two sorting algorithms of the support vector machine of supervised learning, obtain intuitively showing the two-dimensional map figure of high dimensional data, two-category data classification boundaries and data and classification boundaries distance.But directly perceived not in two-dimensional space display, and Various types of data easily overlaps when classification type is more.

These Fault Classifications based on the transmission line of electricity of transient have the advantages that speed is fast, accuracy is high above, in transmission line malfunction classification, obtain good application.

Although the sorting technique of existing document adopts support vector machine can reach good failure modes effect, but when being mingled with noise or bad data in input feature value, the characteristic quantity carried has certain ambiguity and complicacy, different faults type feature amount is not linear separability, the classification accuracy of existing sorting technique will reduce, and even occurs the situation of mis-classification.

List of references:

[1] Chen Shuan, He Zhengyou, LI Xiaopeng. based on the UHV transmission line Fault Phase Selection [J] of traveling wave inherent frequency. electric power network technique, 2011,35 (6): 15-21.

[2] Wang Zhen is tall and erect, Han Fuchun. and the transmission line malfunction based on wavelet transform and fractal theory detects [J]. Institutes Of Technology Of Taiyuan's journal, 2009,40 (3): 300-302.

[3] Hu Dan, Wang Jianhua, Jiang Shupeng, Wang Yating, permitted pine. based on the traveling wave fault phase selection [J] of wavelet transformation. and electric switch, 2013,6:43-45.

[4] Li Xiaochen, Chen Changlei, Zhao Deyang, Li Tianyun. based on the ultra-high-tension power transmission line Fault Phase Selection new method [J] of EMD singular value entropy. China Power, 2011,44 (5): 6-9.

[5] Cui Liyun. based on the transmission line malfunction type identification [J] of EMD Sample Entropy and extreme learning machine. Guangxi electric power, 2012,35 (2): 10-13.

[6] Guo Moufa, Wang Peng, Xu Lilan, Gao Wei, Yang Gengjie. resonant earthed system classification line selection method for single-phase earth fault [P]. Fujian: CN103344875A, 2013-10-09.

[7] superb, Zheng Jianhua, Wang Baohua. based on the transmission line malfunction phase selection new method [J] of SVM. electronic design engineering, 2011,19 (18): 14-17.

[8] Wang Yansong, Tan Zhiyong, Liu Xuemin. based on the distribution network failure type identification [J] of wavelet singular entropy and support vector machine. protecting electrical power system and control, 2011,39 (23): 16-20.

[9] Sun Yaming, Wang Junfeng. based on the new Approach of Fault Type Recognition of Transmission Lines [J] of fractal theory. Automation of Electric Systems, 2005,29 (12): 23-28.

[10] Wu Hao, Luo Yi, Cai Liang. based on the new Approach of Fault Type Recognition of Transmission Lines [J] of RBF neural. Chongqing Mail and Telephones Unvi's journal (natural science edition), 2013,25 (3): 418-426.

[11] WANG Xiaohong, Wang Xiaoru, Lie group is profound. the classification results visualized algorithm [J] of two-category data. and Southwest Jiaotong University's journal, 2006,41 (3): 329-334.

Summary of the invention

The object of this invention is to provide the fuzzy fault sorting technique of the higher transmission line of electricity of a kind of classification accuracy.

Realize technical scheme of the present invention as follows:

A fuzzy fault sorting technique for transmission line of electricity, comprises

Step 1: from the beginning of electric network fault recorded wave file, extracts the three-phase current signal of 4 cycles, carries out EMD decomposition, obtain each intrinsic mode function component IMF, is defined as fault the moment corresponding for IMF1 maximum instantaneous frequency values and the moment occurs;

Step 2: A, B, C three-phase current signal of 2 cycles after obtaining the fault generation moment from failure wave-recording file, these three addition of vectors summations divided by 3, obtains zero sequence current signal; Using A, B, C three-phase current signal and zero sequence current signal as fault sample;

Step 3: A, B, C three-phase current signal of fault sample and zero sequence current signal are carried out EMD decomposition, HHT conversion obtains respective marginal spectrum, select in 0 ~ 2000Hz frequency range, to the integrated square of each marginal spectrum, obtain the feature energy frnction value S of three-phase and zero-sequence current respectively _a, S _b, S _c, S ₀, one 4 dimension input vector of composition, as fault input vector x _g;

Step 4: select radial kernel function RBF as kernel function, fuzzy support vector machine FSVM bis-sorter of structure 10 kinds of fault types, be respectively: A phase short circuit Ag fault FSVM bis-sorter, B phase short circuit Bg fault FSVM bis-sorter, C phase short circuit Cg fault FSVM bis-sorter, B phase C phase short circuit BC fault FSVM bis-sorter, A phase C phase short circuit AC fault FSVM bis-sorter, A phase B phase short circuit AB fault FSVM bis-sorter, A phase B phase ground short circuit ABg fault FSVM bis-sorter, A phase C phase ground short circuit ACg fault FSVM bis-sorter, ground short circuit ABCg fault FSVM bis-sorter when B phase C phase ground short circuit BCg fault FSVM bis-sorter is identical with A phase B phase C,

Step 5: FSVM bis-sorter of each fault type is optimized in training, comprises

Step 5.1: to FSVM bis-sorter of each fault type, the fuzzy training sample set S of each self-structuring one; Described fuzzy training sample set S={ (x ₁, y ₁, u (x ₁)) ... (x _k, y _k, u (x _k)) ... (x _n, y _n, u (x _n)), wherein (x _l, y _l, u (x _l)) be training sample, x _kfor input vector, x _k∈ R ⁿ, _ykfor tag along sort value, y _k{-1,1}, 1 represents faulty tag to ∈, and-1 represents non-faulting label, u (x _k) belong to the fuzzy membership of this fault type, 0≤u (x for training sample _k)≤1;

Step 5.2: to FSVM bis-sorter of each fault type, input the fuzzy training sample set of its correspondence, tries to achieve coefficient vector ω * and the constant term b* of the decision function of this FSVM bis-sorter; Recycling grid optimization algorithm, the punishment parameter C of FSVM bis-sorter of each fault type and kernel function width cs are optimized: to FSVM bis-sorter of each fault type, select the highest one group { C, the σ } of accuracy rate as its optimal classification parameter;

Step 5.3: to FSVM bis-sorter of each fault type, construct the banded segmentation membership function of respective higher dimensional space: obtaining optimal separating hyper plane H by FSVM is ω x+b=0, wherein ω is the normal vector of lineoid, and b is the constant term of lineoid; The mean distance d of positive class sample point to lineoid H is solved in high-dimensional feature space _c, with distance H for d _cand lineoid H parallel with it _cfor reference plane, respectively at H _cboth sides construct different membership function u (x _i),

$u\left(x_i\right)=\left\{\begin{array}{cc}1-\frac{{(d\left(x_i\right)-d_e)}^2}{2.5{d_{\mathrm{ce}}}^2}& (d_e\&le;d\left(x_i\right)\&le;d_c)\\ 0.6-\frac{d\left(x_i\right)}{5d_e}& (0<d\left(x_i\right){}_{}<d_e)\\ 1-\frac{{(d\left(x_i\right)-d_c)}^2}{10{(d_{\max }-d_c)}^2}& (d_c<d\left(x_i\right)\&le;d_{\max })\end{array}\right.$

Wherein, d (x _i) be the distance of arbitrary fault sample point to H, d _ethe distance of support vector to H, d _celineoid H _cto the distance of support vector; d _maxbe the ultimate range of sample point to H, directly give its less degree of membership 0.1 for the negative class sample in negative class side;

Step 5.4: to FSVM bis-sorter of each fault type, input the fuzzy training sample set of its correspondence, tries to achieve the banded segmentation membership function of its higher dimensional space;

Step 6: after the series connection of FSVM bis-sorter of 10 kinds of fault types after optimization, forms combination FSVM sorter of classifying more;

Step 7: by the fault input vector x of fault sample _gbe input in combination many classification FSVM sorter, obtain 10 kinds of fault types preliminary classification label value y separately _g, decision function value Z _gwith initial degree of membership u (x _g);

Step 8: determine that preliminary classification label value yg be that fault type of faulty tag is final fault type.

Further technical scheme is, with following step replacement step 8:

Step 8.1: structure and training support vector regression SVR:

Select radial basis function as kernel function, the support vector regression SVR of structure 10 kinds of fault types;

Each fuzzy training sample set is input to corresponding FSVM bis-sorter of training each fault type after optimization via step 5, obtain the decision function value of the FSVM of each fault type and initial degree of membership, be input to the support vector regression SVR of each corresponding fault type again, try to achieve coefficient vector ω * *, constant term b** after the optimization of this SVR regression function;

Step 8.2: the decision function value Z of 10 kinds of fault types that step 7 is obtained _g, initial degree of membership u (x _g) be input to the support vector regression SVR of corresponding fault type, obtain its regression function value g (x _g) as the final degree of membership of this fault type;

Step 8.3: the maximal value selecting final degree of membership, judges whether it is greater than final degree of membership threshold value g (x ₀):

If so, corresponding preliminary classification label value y is judged _gwhether be faulty tag: be continue; Otherwise by the preliminary classification label value y of correspondence _gbe set to faulty tag, then by original preliminary classification label value y _gnon-faulting label is set to for faulty tag;

If not, continue;

Step 8.4: determine preliminary classification label value y _gthat fault type for faulty tag is final fault type.

Wherein, described final degree of membership threshold value g (x ₀) equal 0.59.

Beneficial effect of the present invention is:

1) the present invention transforms to the characteristic of higher dimensional space linear separability according to sample, and the banded segmentation membership function of higher dimensional space of structure FSVM, obtains the preliminary label of failure modes of FSVM, is under the jurisdiction of the preliminary fuzzy membership of certain fault type.Introduce fuzzy membership function, reduce noise spot, impact that isolated point constructs support vector machine lineoid.

2) the present invention adopts support vector regression SVR, constructs its regression function, optimizes, obtain the coefficient vector ω * * and constant term b** of the regression function of all kinds of fault through training.By test sample book preliminary classification label, be under the jurisdiction of the preliminary fuzzy membership of certain fault type, the final degree of membership of certain class fault is belonged to after obtaining regression optimization by SVR, and can the preliminary classification label obtained by FSVM be revised, thus obtain correct fault type.Like this by the process such as obfuscation, regression optimization, obtain the failure modes label of test sample book exactly, substantially increase accuracy and the fault-tolerance of transmission line malfunction classification.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is the structural representation of the banded segmentation membership function of higher dimensional space;

Fig. 3 is higher dimensional space banded segmentation membership function coordinate diagram;

Fig. 4 is combination many classification FSVM sorter schematic diagram;

Fig. 5 is IEEE14 bus test system schematic diagram.

Embodiment

Principle of the present invention is: introduce the initial fuzzy membership that fuzzy support vector machine obtains sample, then passes through training and the optimization of support vector regression, obtains the final degree of membership of this sample, determines which class fault it belongs to.Utilize at present fuzzy support vector machine and and the research that combines in transmission line malfunction classification of support vector regression still rare at home.

The present invention is first from failure wave-recording file, extract A, B, C three-phase current signal that fault file starts rear 4 cycles, process obtains zero sequence current signal, HHT conversion is adopted to decompose to each signal, obtain each intrinsic mode function (IntrinsicMode Function, IMF) component, the moment corresponding according to the maximum instantaneous frequency of IMF1 is defined as fault and the moment occurs, obtain three-phase current and the zero sequence current signal of 2 cycles after the moment occurs the fault detected, the energy function value S obtaining them, as characteristic quantity, forms each 4 dimension input vectors.For ten kinds of fault types, construct respective fuzzy support vector machine FSVM bis-sorter.Adopt grid optimization method, utilize each training sample that training sample is concentrated, the punishment parameter C of FSVM, kernel function width cs are optimized.Construct banded segmentation membership function, the optimal classification function of FSVM higher dimensional space.The proper vector of test sample book is sent into each FSVM sorter, obtains the tag along sort of certain class fault, calculate the decision function value of this sample and preliminary fuzzy membership.Structure support vector regression SVR and regression function thereof.By to training set sample training and optimization, obtain coefficient vector ω * *, the constant term b** of regression function.The decision function value of this sample and preliminary fuzzy membership are sent into SVR, obtains the final fault degree of membership of certain test sample book.

See Fig. 1, the specific embodiment of the present invention is as follows:

1. after electric network element breaks down, from the beginning of dependent failure recorded wave file, extract A, B, C three-phase current signal of 4 cycles, adopt EMD to decompose, obtain each intrinsic mode function component IMF, is defined as fault the moment corresponding for IMF1 maximum instantaneous frequency values and the moment occurs.

Obtain from failure wave-recording file the three-phase current that 2 cycles after the moment occur for the fault that detects again, these three addition of vectors summations divided by 3, obtain zero sequence current signal.

2. pair A phase, B phase, C phase and zero sequence current signal carry out EMD decomposition, HHT conversion obtains marginal spectrum, select in 0 ~ 2000Hz frequency range, to the integrated square of each marginal spectrum, obtain the feature energy frnction value S of three-phase and zero-sequence current respectively _a, S _b, S _c, S ₀, form one 4 dimension input vector.Computing method are as follows.

The analytical expression of original signal is obtained such as formula shown in (1) through EMD decomposition and HHT conversion.

$Z_k\left(t\right)={\mathrm{imf}}_k\left(t\right)+{\mathrm{jh}}_k\left(t\right)=a_k\left(t\right)e^{j{\mathrm{\&theta;}}_k\left(t\right)}---\left(1\right)$

In formula (1):

$h_k\left(t\right)=\frac{1}{\mathrm{\&pi;}}P{\&Integral;}_{-\&infin;}^{+\&infin;}\frac{{\mathrm{imf}}_k\left(\mathrm{\&delta;}\right)}{t-\mathrm{\&delta;}}\mathrm{d\&delta;}---\left(2\right)$

Calculate Hilbert spectrum:

$H(\mathrm{\&omega;},t)=\mathrm{Re}\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_k\left(t\right)e^{j\&Integral;\mathrm{\&omega;}\left(t\right)d\left(t\right)}$

Again at [0, T], integration is carried out to Hilbert spectrum, obtain Hilbert marginal spectrum, shown in (4).

$h\left(\mathrm{\&omega;}\right)={\&Integral;}_0^{T}H(\mathrm{\&omega;},t)\mathrm{dt}---\left(4\right)$

In certain band limits, square integration is carried out to marginal spectrum, obtain feature energy frnction value S.

$S={\&Integral;}_{{\mathrm{\&omega;}}_1}^{{\mathrm{\&omega;}}_2}{h}^2\left(\mathrm{\&omega;}\right)d\left(\mathrm{\&omega;}\right)---\left(5\right)$

Select 0 ~ 2000Hz frequency range, calculate the feature energy frnction value S of the A phase of each sample, B phase, C phase, zero sequence current signal respectively _a, S _b, S _c, S ₀, formed one 4 dimension input vector x with these 4 components _g.

3. construct the method for FSVM and decision function thereof.

If input fuzzy training set S:

S＝{(x ₁,y ₁,u(x ₁)),...(x _k,y _k,u(x _k))...(x _n,y _n,u(x _n))}

Wherein (x _k, y _k, u (x _k)) be a training sample, x _kfor input vector, x _k∈ R ⁿ, y _kfor tag along sort value, y _k∈-1,1} (1 represent belong to this type of ,-1 represent non-this type of), u (x _k) belong to the fuzzy membership of certain fault type, 0 for training sample ^≤u (x _k) ^≤1.

The structure of FSVM optimal separating hyper plane can be summed up as Quadratic Programming Solution, in this quadratic programming, introduce fuzzy membership, and the optimal separating hyper plane of FSVM is expressed as:

$\begin{array}{c}\underset{\mathrm{\&omega;},b,\mathrm{\&xi;}}{\min }\frac{1}{2}{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}\mathrm{\&mu;}\left(p_i\right){\mathrm{\&xi;}}_i\\ s.t.y_i((\mathrm{\&omega;}\&CenterDot;p_i)+b)+{\mathrm{\&xi;}}_i\&GreaterEqual;1\\ {\mathrm{\&xi;}}_i\&GreaterEqual;0,i=\mathrm{1,2},...l\end{array}---\left(6\right)$

In formula, C > 0 is punishment parameter, relaxation factor ξ=(ξ ₁, ξ ₂... ξ _i) ^t, μ (p _i) ξ _ibe the wrong point degree of sample Measure Indexes (i=1 ..., l).

The objective function that the dual program obtaining formula (6) solves:

$\begin{array}{c}\underset{\mathrm{\&alpha;}}{\min }\frac{1}{2}\underset{j=1}{\overset{l}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{y}_j{y}_i{\mathrm{\&alpha;}}_j{\mathrm{\&alpha;}}_{i}K(p_i,p_j)-\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i\\ s.t.\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_i=0\\ 0\&le;{\mathrm{\&alpha;}}_i\&le;\mathrm{\&mu;}\left(p_i\right)C,i=1,...,l\end{array}---\left(7\right)$

Wherein K (p _i, p _j) be kernel function, select RBF as kernel function.Obtain the optimal decision function of FSVM:

$f\left(x_j\right)=\underset{s=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}^{*}K(x_j,p_s)+b^{*},x_j\&Element;R^n---\left(8\right)$

ω in formula (8) ^*=α _i ^*y _ifor coefficient vector, b ^*for constant term vector, x _jfor the sample vector of input, p _sfor support vector.

Each training sample vector substitution formula (8), the coefficient vector ω of the FSVM decision function of all kinds of fault can be obtained ^*, constant term b ^*.

For certain training sample, substitute into formula (8), corresponding decision function value Z can be obtained _k.

For fault sample x _g, obtain by formula (9) the preliminary classification label functional value y that it belongs to certain class fault _g=f (x _g).

$f\left(x_g\right)=\mathrm{sgn}\underset{s=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}^{*}\{K(x_g,p_s)+b^{*}\},x_g\&Element;R^n---\left(9\right)$

4. based on grid optimization algorithm (the Gridding Optimization Algorithm of the node searching of grid, GOA) be a kind of direction-sense global search method, utilize the punishment parameter C kernel function width cs of this grid optimization algorithm to all kinds of fault FSVM model to be optimized.Select one group { C, σ } that accuracy rate is the highest, as the optimal classification parameter of certain failure modes FSVM.

5. construct the banded segmentation membership function of FSVM higher dimensional space.Building method is as follows:

Obtaining optimal separating hyper plane H by FSVM is ω x+b=0, solves the mean distance d of positive class sample point to lineoid H in high-dimensional feature space _c, with distance H for d _cand lineoid H parallel with it _cfor reference plane, respectively at H _cboth sides construct different membership functions.The structural representation of the banded segmentation membership function of higher dimensional space as shown in Figure 2.

Higher dimensional space banded segmentation membership function u (x _i) design as follows.

$u\left(x_i\right)=\left\{\begin{array}{cc}1-\frac{{(d\left(x_i\right)-d_e)}^2}{2.5{d_{\mathrm{ce}}}^2}& (d_e\&le;d\left(x_i\right)\&le;d_c)\\ 0.6-\frac{d\left(x_i\right)}{5d_e}& (0<d\left(x_i\right)<d_e)\\ 1-\frac{{(d\left(x_i\right)-d_c)}^2}{10{(d_{\max }-d_c)}^2}& (d_c<d\left(x_i\right)\&le;d_{\max })\end{array}\right.---\left(10\right)$

D (x in formula (10) _i) be the distance of arbitrary sample point to H, d _ethe distance of support vector to H, d _celineoid H _cto the distance of support vector, d _maxthe ultimate range of sample point to H.Its less degree of membership 0.1 is directly given for the negative class sample in negative class side.According to the distance of each sample to lineoid H both sides, carry out the initial degree of membership u (x that Fuzzy Processing asks for training sample _i).Higher dimensional space banded segmentation membership function coordinate diagram as shown in Figure 3.

6. according to the method that above 3-5 walks, construct two sorters of all types of fault of transmission line of electricity respectively, combination many classification FSVM sorter schematic diagram as shown in Figure 4.The structure of each two sorters comprises the coefficient vector ω of their decision function ^*, constant term b ^*.Calculate the preliminary classification label value y of fault sample _g, decision function value Z _g.The initial degree of membership u (x of this sample is calculated by formula (10) _g).By preliminary classification label value y _gfault type can be determined.

Further, following process is also comprised:

7. construct the regression function of the support vector regression SVR of two sorters of all types of fault, construct its optimum regression function g (x).Method is as follows:

$g\left(x\right)=\underset{q=1}{\overset{l}{\mathrm{\&Sigma;}}}\mathrm{\&omega;K}(x,x_q)+b,x\&Element;R^n---\left(11\right)$

The objective function of regression function is represented by formula (12).

$\begin{array}{c}\min \frac{1}{2}{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i\\ \begin{array}{cc}s.t.& \mathrm{\&mu;}\left(t_i\right)-(\mathrm{\&omega;}\&CenterDot;t_i)-b\&le;\mathrm{\&epsiv;}+{\mathrm{\&xi;}}_i\end{array}\\ \begin{array}{cc}s.t.& (\mathrm{\&omega;}\&CenterDot;t_i)+b-\mathrm{\&mu;}\left(t_i\right)\&le;\mathrm{\&epsiv;}+{\mathrm{\&xi;}}_i\end{array}\\ \begin{array}{cc}i=1,...n& {\mathrm{\&xi;}}_i\&GreaterEqual;0\end{array}\end{array}---\left(12\right)$

The regression function of support vector regression SVR is:

$g\left(t_i\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)K(t_i,t_j)+b---\left(13\right)$

α in formula (13) ^*it is optimum solution.

Select radial basis function as the kernel function of SVR.

The input vector { Z of training set sample _k, u (x _k) (Z _kfor the decision function value that input vector is corresponding, u (x _k) be the initial degree of membership that this sample belongs to certain class fault) substitute into formula (13) and train, can obtain regression function optimize after coefficient vector ω ^*, constant term b ^*.

The input vector x of fault sample _gsubstitution formula (12), can obtain its regression function value g (x _g), namely obtain the final degree of membership that this sample belongs to certain class fault.

8. set fault threshold, the fault type of judgement sample.

The present invention sets fault threshold g (x ₀) be 0.59.Select the maximal value of final degree of membership, judge whether it is greater than final degree of membership threshold value g (x ₀).If the maximal value of final degree of membership is greater than threshold value g (x ₀), so judge corresponding preliminary classification label value y _gwhether be faulty tag, be, determine that this fault type is final fault type.If the maximal value of final degree of membership is greater than threshold value g (x ₀), corresponding preliminary classification label value y _gnot faulty tag, so by the preliminary classification label value y of correspondence _gbe set to faulty tag, then will originally walk tag along sort value y _gnon-faulting label is set to for faulty tag.If the maximal value of final degree of membership is not more than threshold value g (x ₀), so preliminary classification label value is not made an amendment.Finally, preliminary classification label value y is determined _gthat fault type for faulty tag is final fault type.

Fig. 5 is IEEE14 bus test system.IEEE14 node realistic model emulates, selects different faults type, trouble spot, transition resistance, obtain all kinds of fault-current signal.If the long 180km of circuit 15, different faults point, transition resistance is chosen respectively then at node 9 place on this circuit, obtain and change by trouble spot, transition resistance each three-phase current obtained, calculate corresponding zero sequence current signal, HHT conversion is carried out to these current signals, obtains the feature energy frnction value S of 0 ~ 2000Hz frequency range.

For all kinds of fault type, at circuit 15, trouble spot (totally 19 positions) is set every total length 5%, 10 kinds of fault types, transition resistance establishes 0,50,100,200 Ω (totally 4 kinds of situations) respectively, add that electrical network normally runs 2 groups of data, altogether be provided with 19 × 4 × 10+2=782 group data, as training set sample.Extract A, B, C three-phase and zero-sequence current characteristic quantity often organized, form each 4 dimensional feature vectors, as the input vector of FSVM.

In addition, distance 9 node on circuit 15 is selected to be respectively total track length 20%, 40%, 60%, 80% trouble spot (totally 4 positions), 10 kinds of fault types, transition resistance is respectively 25,75,125,175 Ω (totally 4 kinds of situations), in addition current data 1 group when electrical network normally runs, has prepared 4 × 10 × 4+1=161 group data altogether as test set sample.

By training the FSVM sorter of often kind of fault type, obtain the parameter vector of often kind of FSVM sorter.Send into each test set sample, obtain the classification results of SVM, FSVM and FSVM+SVR tri-kinds of sorting techniques respectively.

Then after adding 5dB noise in training set 1/5 sample, often kind of FSVM sorter is trained respectively, then test experiments is carried out to SVM, FSVM and FSVM+SVR tri-kinds of sorting techniques, obtain corresponding classification results.

1.SVM, FSVM, FSVM+SVR classification results.

The discrimination statistics of table 1 SVM, FSVM, FSVM+SVR classification results

2. training set fault phase 1/5 data add the classification results of SVM, FSVM, FSVM+SVR after 5dB noise.

SVM, FSVM, FSVM+SVR classification results discrimination contrast after table 2 plus noise

Claims (3)

1. a fuzzy fault sorting technique for transmission line of electricity, is characterized in that, comprise

Step 1: from the beginning of electric network fault recorded wave file, extracts the three-phase current signal of 4 cycles, carries out EMD decomposition, obtain each intrinsic mode function component IMF, is defined as fault the moment corresponding for IMF1 maximum instantaneous frequency values and the moment occurs;

Step 2: A, B, C three-phase current signal of 2 cycles after obtaining the fault generation moment from failure wave-recording file, these three addition of vectors summations divided by 3, obtains zero sequence current signal; Using A, B, C three-phase current signal and zero sequence current signal as fault sample;

Step 3: A, B, C three-phase current signal of fault sample and zero sequence current signal are carried out EMD decomposition, HHT conversion obtains respective marginal spectrum, select in 0 ~ 2000Hz frequency range, to the integrated square of each marginal spectrum, obtain the feature energy frnction value S of three-phase and zero-sequence current respectively _a, S _b, S _c, S ₀, one 4 dimension input vector of composition, as fault input vector x _g;

Step 4: select radial kernel function RBF as kernel function, fuzzy support vector machine FSVM bis-sorter of structure 10 kinds of fault types, be respectively: A phase short circuit Ag fault FSVM bis-sorter, B phase short circuit Bg fault FSVM bis-sorter, C phase short circuit Cg fault FSVM bis-sorter, B phase C phase short circuit BC fault FSVM bis-sorter, A phase C phase short circuit AC fault FSVM bis-sorter, A phase B phase short circuit AB fault FSVM bis-sorter, A phase B phase ground short circuit ABg fault FSVM bis-sorter, A phase C phase ground short circuit ACg fault FSVM bis-sorter, ground short circuit ABCg fault FSVM bis-sorter when B phase C phase ground short circuit BCg fault FSVM bis-sorter is identical with A phase B phase C,

Step 5: FSVM bis-sorter of each fault type is optimized in training, comprises

Step 5.1: to FSVM bis-sorter of each fault type, the fuzzy training sample set S of each self-structuring one; Described fuzzy training sample set S={ (x ₁, y ₁, u (x ₁)) ... (x _k, y _k, u (x _k)) ... (x _n, y _n, u (x _n)), wherein (x _l, y _l, u (x _l)) be training sample, x _kfor input vector, x _k∈ R ⁿ, y _kfor tag along sort value, y _k{-1,1}, 1 represents faulty tag to ∈, and-1 represents non-faulting label, u (x _k) belong to the fuzzy membership of this fault type, 0≤u (x for training sample _k)≤1;

Step 5.2: to FSVM bis-sorter of each fault type, input the fuzzy training sample set of its correspondence, try to achieve the coefficient vector ω of the decision function of this FSVM bis-sorter ^*with constant term b ^*; Recycling grid optimization algorithm, the punishment parameter C of FSVM bis-sorter of each fault type and kernel function width cs are optimized: to FSVM bis-sorter of each fault type, select the highest one group { C, the σ } of accuracy rate as its optimal classification parameter;

Step 5.3: to FSVM bis-sorter of each fault type, construct the banded segmentation membership function of respective higher dimensional space: obtaining optimal separating hyper plane H by FSVM is ω x+b=0, wherein ω is the normal vector of lineoid, and b is the constant term of lineoid; The mean distance d of positive class sample point to lineoid H is solved in high-dimensional feature space _c, with distance H for d _cand lineoid H parallel with it _cfor reference plane, respectively at H _cboth sides construct different membership function u (x _i),

$u\left(x_i\right)=\left\{\begin{array}{cc}1-\frac{{(d\left(x_i\right)-d_e)}^2}{2.5{d_{\mathrm{ce}}}^2}& (d_e\&le;d\left(x_i\right)\&le;d_c)\\ 0.6-\frac{d\left(x_i\right)}{5d_e}& (0<d\left(x_i\right)<d_e)\\ 1-\frac{{(d\left(x_i\right)-d_c)}^2}{10{(d_{\max }-d_c)}^2}& (d_c<d\left(x_i\right)\&le;d_{\max })\end{array}\right.$

Wherein, d (x _i) be the distance of arbitrary sample point to H, d _ethe distance of support vector to H, d _celineoid H _cto the distance of support vector; d _maxbe the ultimate range of sample point to H, directly give its less degree of membership 0.1 for the negative class sample in negative class side;

Step 5.4: to FSVM bis-sorter of each fault type, input the fuzzy training sample set of its correspondence, tries to achieve the banded segmentation membership function of its higher dimensional space;

Step 6: after the series connection of FSVM bis-sorter of 10 kinds of fault types after optimization, forms combination FSVM sorter of classifying more;

Step 7: by the fault input vector x of fault sample _gbe input in combination many classification FSVM sorter, obtain 10 kinds of fault types preliminary classification label value y separately _g, decision function value Z _gwith initial degree of membership u (x _g);

Step 8: determine preliminary classification label value y _gthat fault type for faulty tag is final fault type.

2. the fuzzy fault sorting technique of transmission line of electricity as claimed in claim 1, is characterized in that, with following step replacement step 8:

Step 8.1: structure and training support vector regression SVR:

Select radial basis function as kernel function, the support vector regression SVR of structure 10 kinds of fault types;

Each fuzzy training sample set is input to corresponding FSVM bis-sorter of training each fault type after optimization via step 5, obtain the decision function value of the FSVM of each fault type and initial degree of membership, be input to the support vector regression SVR of each corresponding fault type again, try to achieve the coefficient vector ω after the optimization of this SVR regression function ^*, constant term b ^*;

Step 8.2: the decision function value Z of 10 kinds of fault types that step 7 is obtained _g, initial degree of membership u (x _g) be input to the support vector regression SVR of corresponding fault type, obtain its regression function value g (x _g) as the final degree of membership of this fault type;

Step 8.3: the maximal value selecting final degree of membership, judges whether it is greater than final degree of membership threshold value g (x ₀):

If so, corresponding preliminary classification label value y is judged _gwhether be faulty tag: be continue; Otherwise preliminary by correspondence

Tag along sort value y _gbe set to faulty tag, then by original preliminary classification label value y _gnon-faulting label is set to for faulty tag;

If not, continue;

Step 8.4: determine preliminary classification label value y _gthat fault type for faulty tag is final fault type.

3. the fuzzy fault sorting technique of transmission line of electricity as claimed in claim 2, is characterized in that, described final degree of membership threshold value g (x ₀) equal 0.59.